lagent

Lagent is a lightweight framework designed to facilitate the creation of LLM-based agents. Inspired by PyTorch, Lagent aims to provide a clear and intuitive workflow for building multi-agent applications.

Key Features:

• Agent Communication: Utilizes AgentMessage for seamless communication between agents.
• Memory Management: Incorporates memory within agents to maintain state across interactions. Input and output messages are automatically added to the agent's memory.
• Customizable Aggregation: Offers DefaultAggregator for converting AgentMessage to OpenAI message format, with the option to implement custom aggregators for specific use cases.
• Flexible Response Formatting: Supports flexible response formatting through output parsers, allowing structured extraction of information from model outputs.
• Tool Calling Consistency: Ensures consistency in tool calling through ActionExecutor, which requires input AgentMessage to contain tool name and parameters.
• Dual Interfaces: Provides both synchronous and asynchronous interfaces for LLMs, actions, and action executors, enabling optimized performance for debugging and large-scale inference.

Use Cases:

• Math Agents: Solving mathematical problems using Python code execution.
• Blogging Agents: Improving writing quality through self-refinement workflows.
• Data Visualization Agents: Performing information retrieval, data collection, and chart plotting for data analysis and presentation.